A method for predicting cell network load indicators based on attention mechanism learning network

Abstract

An embodiment of the present invention provides a method for predicting cell network load indicators based on an attention mechanism learning network. The method includes collecting historical data of the cell to be predicted to obtain the original time series to be predicted, then performing preset rule processing to obtain the clean time series to be predicted; inputting the clean time series to be predicted into a prediction model, and outputting the predicted mobile communication network load indicator for the cell on that day; wherein the prediction model is obtained by training based on sample clean time series of M cells and the mobile communication network load indicator labels corresponding to the M cells on that day, and the network structure used in training the prediction model is a BiGRU network with an attention mechanism added. The method provided by the embodiment of the present invention improves the accuracy of cell mobile communication network load indicator prediction.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method for predicting cell network load indicators based on an attention mechanism learning network. Background Art

[0002] With the development of mobile wireless networks, especially the continuous evolution of 4G and 5G networks, mobile networks are facing increasingly serious load problems. The regional and temporal uneven distribution of network load is becoming increasingly prominent. How to achieve accurate and rapid prediction of mobile communication network load indicators is an important prerequisite for ensuring the capacity of mobile wireless networks.

[0003] Among existing technical solutions, the most commonly used method is modeling through LSTM networks. First, load data for a certain historical time period is obtained, then normalized. The normalized data is input into the LSTM network for training, and the prediction model is constructed through cyclic iteration. The existing communication capability open load forecasting method based on multi-layer LSTM-RNN obtains load data from the communication capability open platform within a preset time period before the time period to be predicted, normalizes the load data, and transforms the normalized load data into a shape suitable for the LSTM neural network. The load data includes the hourly API call volume and hourly peak TPS at an hourly granularity. The transformed load data is input into a pre-built and trained multi-layer LSTM-RNN load forecasting model, and the output result is denormalized to obtain the load forecast result for the time period to be predicted. The multi-layer LSTM-RNN load forecasting model includes an input layer, an output layer, and at least two stacked LSTM hidden layers. However, the solution using LSTM network for modeling and prediction has a simple network structure. The input gate and forget gate in the network are complementary and have a certain degree of redundancy. It fails to introduce relevant corrections and data preprocessing based on the characteristics of mobile communication network load indicators, resulting in poor prediction results.

[0004] Another type of method in the prior art uses linear models to predict load index data. Based on the load variation characteristics of a specific scenario, a linear model is established to predict load data for similar scenarios. Among these methods, existing mobile communication network traffic prediction methods obtain the paging volume of the first mobile communication network based on paging information from the first mobile communication network, and obtain the traffic volume of the first mobile communication network based on the traffic information of the first mobile communication network. A linear regression model is then developed based on the paging volume and traffic volume of the first mobile communication network using the principle of linear regression. The paging volume of the second mobile communication network is obtained based on the paging information of the second mobile communication network. The first and second mobile communication networks are different, and the traffic volume of the second mobile communication network is estimated based on the paging volume of the second mobile communication network and the linear relationship model between the paging volume and traffic volume. These linear modeling prediction methods require the establishment of separate models for different scenarios, and linear models can only be applied to a few specific scenarios, resulting in poor adaptability. The load index of a mobile communication network varies complexly and is influenced by numerous factors, making linear models inaccurate.

[0005] Therefore, how to avoid the problem of low scenario adaptability caused by low accuracy of mobile communication network load index prediction in the existing technology, and how to make the mobile communication network load index prediction meet the requirements of high-precision prediction, are still problems that technical personnel in this field need to solve urgently. Summary of the Invention

[0006] The embodiments of the present invention provide a method and device for predicting cell network load indicators based on an attention mechanism learning network, so as to solve the problem of low accuracy in predicting mobile communication network load indicators in the prior art.

[0007] In a first aspect, an embodiment of the present invention provides a method for predicting a cell network load index based on an attention mechanism learning network, comprising:

[0008] Collecting historical data of the mobile communication network load index to be predicted for the cell area of ​​the mobile communication network load index to be predicted for the previous N days to obtain an original time series to be predicted, and processing the original time series to be predicted according to a preset rule to obtain a clean time series to be predicted, where N is a positive integer;

[0009] Inputting the cleaning time series to be predicted into the prediction model, and outputting the predicted mobile communication network load index of the cell on that day;

[0010] Among them, the prediction model is obtained after training based on the sample cleaning time series of M cells and the mobile communication network load index labels corresponding to the M cells on the same day. The sample cleaning time series of the M cells is obtained by processing the sample original time series obtained by collecting the sample historical data of the mobile communication network load index of the M cells in the previous N days and performing the preset rule processing. The network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0011] Preferably, in the method, the step of processing the original time series to be predicted according to preset rules to obtain the cleaning time series to be predicted specifically includes:

[0012] Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted;

[0013] Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted;

[0014] Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted;

[0015] Normalizing and standardizing the to-be-predicted completion time series to obtain a to-be-predicted cleaning time series;

[0016] The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by training the random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells.

[0017] Preferably, in this method, the outlier detection model is obtained using a One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, specifically comprising:

[0018] The objective function and constraints of the following optimization problem are constructed to solve the outlier detection model using the One-Class SVM outlier detection algorithm:

[0019]

[0020]

[0021]

[0022] Where a and R are the center and radius of the hypersphere constructed based on n normal sample logarithmic data, R>0, C is the preset penalty coefficient, is the preset slack variable, x i Represents the distance from the i-th mobile communication network load index data point to the center. The sample logarithmic time series of the first N days of M cells contains n normal sample logarithmic data.

[0023] Preferably, in this method, the completion model is obtained by training a random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells, specifically comprising:

[0024] For the tth iteration of training,

[0025] Fitting and Get the fitting matrix and train the current iteration round random forest model Rf t ,Will Input Rf t , output

[0026] use Update the fitting matrix, s=1, 2, ..., k;

[0027] Where t = 1, 2, …, γ, γ is the preset maximum iteration threshold, k is the number of cells with missing data in the normal logarithmic time series of samples of the first N days of M cells, is the observed value of the sample normal logarithmic time series of the sth cell with missing data, Completion prediction of missing values ​​for the normal logarithmic time series of samples in the sth cell with missing data, is the observed value of the normal logarithmic time series of samples of M cells except the missing data corresponding to the s-th cell in the normal logarithmic time series of other samples except the s-th cell, is the observed value of the missing data corresponding to the sth cell in the normal logarithmic time series of samples of the M cells except the sth cell;

[0028] The random forest model Rfγ obtained by the γth iterative training is the completion model.

[0029] Preferably, in this method, the network structure used in the prediction model training is a BiGRU network with an Attention mechanism added, specifically including:

[0030] The prediction model is trained using a 1-layer BiGRU network layer and the output h of the 1-layer BiGRU network layer is converted into l And the output Y of all BiGRU network layers is input into the Attention mechanism layer;

[0031] The Attention mechanism layer performs the following operations:

[0032] Based on h l and Y determine the attention feature score of each BiGRU network layer at each moment;

[0033] Determine the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment;

[0034] Determine a weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers;

[0035] Based on the weighted expression of all BiGRU network layer outputs and h l Determine the output of the Attention mechanism layer.

[0036] Preferably, in the method, the h-based l and Y determine the attention feature scores of each BiGRU network layer at each moment, including:

[0037] The attention feature score MQ of each BiGRU network layer at each moment is determined by the following formula:

[0038] MQ=tanl(W y ×Y+(W h ×h l ×e l ))

[0039] Among them, l is the number of BiGRU network layers, h l ∈R 1×l The first layer is the output of the BiGRU network layer, MQ∈R F×l , F is the number of elements in the cleaning time series of each cell sample, W y ∈R F×F , Wh∈R F×F , e l ∈R F And W y , Wh and e l are randomly generated trainable matrices, matrices, and vectors, Y∈R F×l And it is the output of all BiGRU network layers;

[0040] Correspondingly, determining the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment specifically includes:

[0041] The weight value aa corresponding to each BiGRU network layer at the current moment is determined by the following formula:

[0042] aa=softmax(ωT ×MQ)

[0043] Where aa∈R l ,ω∈R F And ω is a randomly generated trainable vector, and softmax(·) is a normalized exponential function;

[0044] The determining of the weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers specifically includes:

[0045] The weighted expression rr of all BiGRU network layer outputs is determined by the following formula:

[0046] rr=Yaa T

[0047] Where rr∈R F ;

[0048] The weighted expression based on the output of all BiGRU network layers and h l Determine the output of the Attention mechanism layer, including:

[0049] The output h of the Attention mechanism layer is determined by the following formula * :

[0050] h * =tanl(W p ×rr+W x ×h l )

[0051] Among them, W p ∈R F×F , W x ∈R F×F , and W p and W x All are randomly generated trainable matrices.

[0052] In a second aspect, an embodiment of the present invention provides a cell network load index prediction device based on an attention mechanism learning network, comprising:

[0053] a preprocessing unit, configured to collect historical data of the mobile communication network load index to be predicted for the previous N days of the cell of the mobile communication network load index to be predicted to obtain an original time series to be predicted, and process the original time series to be predicted according to a preset rule to obtain a clean time series to be predicted, wherein N is a positive integer;

[0054] A prediction unit is used to input the cleaning time series to be predicted into a prediction model and output the predicted mobile communication network load index of the cell on that day; wherein, the prediction model is obtained after training based on the sample cleaning time series of M cells and the mobile communication network load index labels corresponding to the M cells on that day, and the sample cleaning time series of the M cells are obtained by processing the sample original time series obtained by collecting the sample historical data of the mobile communication network load indicators of the M cells in the previous N days according to the preset rules; the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0055] Preferably, in the device, the step of processing the original time series to be predicted according to preset rules to obtain the cleaning time series to be predicted specifically includes:

[0056] Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted;

[0057] Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted;

[0058] Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted;

[0059] Normalizing and standardizing the to-be-predicted completion time series to obtain a to-be-predicted cleaning time series;

[0060] The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by training the random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells.

[0061] In a third aspect, an embodiment of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for predicting a cell network load index based on an attention mechanism learning network as provided in the first aspect are implemented.

[0062] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the cell network load index prediction method based on the attention mechanism learning network provided in the first aspect.

[0063] The embodiment of the present invention provides a method and device for predicting a cell network load index based on an attention mechanism learning network, which collects historical data of the mobile communication network load index of the cell to be predicted in the previous N days to obtain an original time series to be predicted, and processes the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; the clean time series to be predicted is input into a prediction model, and the predicted mobile communication network load index of the cell on that day is output; wherein the prediction model is obtained after training based on sample clean time series of M cells and mobile communication network load index labels of the M cells on that day, and the sample clean time series of the M cells is obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load indicators of the M cells in the previous N days according to the preset rules; the network structure used in training the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer. Because the prediction model for mobile communication network load index prediction is dynamically generated through deep learning, and the BiGRU network used for model training incorporates an attention mechanism, which is used during training to screen important features and filter out low-value features, the prediction model's ability to express mobile communication network load index data collected over a long period of time is improved. Through iterative training, the model can more accurately fit the changing trends of mobile communication network load index data. Therefore, the method and apparatus provided by the embodiments of the present invention achieve improved accuracy in predicting cell mobile communication network load indexes. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 A schematic diagram of a flow chart of a method for predicting cell network load indicators based on an attention mechanism learning network provided by an embodiment of the present invention;

[0066] Figure 2 A schematic diagram of a network structure in which a multi-layer BiGRU network is superimposed according to an embodiment of the present invention;

[0067] Figure 3 A schematic diagram of the structure of a network model with a multi-layer BiGRU plus Attention mechanism provided by an embodiment of the present invention;

[0068] Figure 4A schematic diagram of the structure of a cell network load index prediction device for an attention mechanism learning network provided by an embodiment of the present invention;

[0069] Figure 5 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] The existing methods for predicting mobile communication network load indicators generally have the problem of inaccurate prediction results. To address this issue, an embodiment of the present invention provides a method for predicting cell network load indicators based on an attention mechanism learning network. Figure 1 A flowchart of a method for predicting cell network load indicators based on an attention mechanism learning network provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0072] Step 110 , collecting historical data of the mobile communication network load index to be predicted for the cell in which the mobile communication network load index is to be predicted for the previous N days to obtain an original time series to be predicted, and subjecting the original time series to be predicted to preset rule processing to obtain a clean time series to be predicted, wherein N is a positive integer.

[0073] Specifically, the cell network load index prediction method provided by the embodiment of the present invention is based on the historical data of the cell whose mobile communication network load index is to be predicted, which is then pre-processed and then input into the prediction model. The prediction model outputs the predicted data. The historical data of the cell whose mobile communication network load index is to be predicted is collected by collecting the historical data of the mobile communication network load index for the previous N days before the load index of the current day is to be predicted. For example, N = 30, that is, collecting the historical data of the previous month. At the same time, the historical data can also be collected at intervals of unit time, for example, every hour, that is, collecting 24 data per day. Therefore, the historical data to be predicted for the previous 30 days includes 30×24 data. These 30×24 data are then arranged in chronological order to obtain the original time series to be predicted. Here, N is a positive integer. Considering the balance between prediction accuracy and computational complexity, N is usually taken as 30. For the daily collection method of historical data to be predicted, it is preferably collected every hour, that is, 24 cell network load index data are collected in one day. The prediction is made through the historical data collected every hour, and the output prediction data for the day is also the value of the cell network load index corresponding to each hour of the 24 hours of the day. Usually, the cell network load index data can be represented by network traffic, PRB utilization, and the number of RRC users. Preferably, the cell network load index is the size of the network traffic. At the same time, the original time series to be predicted must also be processed according to preset rules. The preset rule processing is to clean the data so that the data input into the prediction model is guaranteed to be free of anomalies, complete, and convenient for model calculation. Usually, the preset rules include data format conversion, outlier removal, missing data completion, normalization, and standardization steps.

[0074] Step 120: input the cleaning time series to be predicted into the prediction model, and output the predicted mobile communication network load index of the cell on that day; wherein, the prediction model is obtained after training based on the sample cleaning time series of M cells and the mobile communication network load index labels corresponding to the M cells on that day, and the sample cleaning time series of the M cells are obtained by processing the sample original time series obtained by collecting the sample historical data of the mobile communication network load index of the M cells in the previous N days according to the preset rules; the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0075] Specifically, the clean time series to be predicted obtained after processing in step 110 is input into the prediction model, and the predicted mobile communication network load index of the cell on that day is output. The prediction model is obtained after training based on the sample clean time series of M cells and the mobile communication network load index labels corresponding to the M cells on that day, wherein the sample clean time series of M cells used for model training is obtained by processing the sample original time series obtained by collecting the sample historical data of the mobile communication network load indicators of the M cells in the previous N days using the same preprocessing method as the preset rules used in step 110, that is, if the preset rules in step 110 include data format conversion, outlier removal, missing data filling, normalization and standardization steps, then the sample original time series obtained by collecting the sample historical data of the mobile communication network load indicators of the M cells in the previous N days in step 120 is also processed through the same data format conversion, outlier removal, missing data filling, normalization and standardization steps to obtain the sample clean time series of M cells used for model training. It should be noted that the number of elements in the sample clean time series used for training and the predicted clean time series input during model use are identical and correspond one-to-one, meaning that each element corresponds to the same time. For example, if the sample clean time series used for model training has a dimension of 30×24, representing the indicator data collected for the sample cell for the previous 30 days, then the input clean time series for prediction also contains the same 30×24 elements. These 30×24 elements are also composed of the indicator data for the cell for the previous 30 days, arranged in chronological order. The prediction results output by the model are also the predicted indicator values ​​for the 24 hours of the current day. Furthermore, during the training process of the prediction model, a multi-layer BiGRU network is used to construct the training network. Attention is also incorporated into this multi-layer BiGRU network to select important features and filter out low-value features, improving the prediction model's ability to express the cell's long-term mobile communication network load indicators. Through iterative training, it can more accurately fit the changing trends of network data.

[0076] The method provided by an embodiment of the present invention collects historical data of the mobile communication network load index to be predicted for the previous N days of the cell whose mobile communication network load index is to be predicted to obtain an original time series to be predicted, and processes the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; the clean time series to be predicted is input into a prediction model, and the predicted mobile communication network load index of the cell on that day is output; wherein the prediction model is obtained after training based on sample clean time series of M cells and mobile communication network load index labels of the M cells on that day, and the sample clean time series of the M cells is obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load index of the M cells in the previous N days according to the preset rules, and the network structure used in training the prediction model is a BiGRU network with an Attention mechanism added. Because the prediction model for mobile communication network load index prediction is dynamically generated through deep learning, and the BiGRU network used for model training incorporates an attention mechanism, which is used during training to screen important features and filter out low-value features, the prediction model's ability to express mobile communication network load index data collected over a longer period of time is improved. Through iterative training, the model can more accurately fit the changing trends of mobile communication network load index data. Therefore, the method provided by the embodiments of the present invention achieves improved accuracy in predicting cell mobile communication network load indexes.

[0077] Based on the above embodiment, in this method, the step of processing the original time series to be predicted according to a preset rule to obtain the cleaning time series to be predicted specifically includes:

[0078] Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted;

[0079] Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted;

[0080] Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted;

[0081] Normalizing and standardizing the to-be-predicted completion time series to obtain a to-be-predicted cleaning time series;

[0082] The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by training the random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells.

[0083] Specifically, the preset rule processing of the original time series to be predicted includes the five steps of format conversion, outlier removal, missing data completion, normalization and standardization.

[0084] Among them, format conversion is to convert the collected data from the natural number domain to the logarithmic domain, that is, to take the decimal logarithm of the original network load index data, obtain the logarithmic value, and form the logarithmic time series to be predicted;

[0085] The outlier detection model is obtained using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days in M ​​communities;

[0086] The completion model is trained using a random forest classifier based on the normal logarithmic time series of samples from the first N days of M cells;

[0087] After the abnormal points are eliminated and the missing data are filled, the time series to be predicted is normalized. For example, the time series to be predicted formed by the previous N days of historical data of the cell to be predicted for the load index of the mobile communication network is X, where X is a column vector with a dimension of m, X = {x1, x2, ..., x m}, m = N × Q, where Q is the number of data that needs to be collected every day when collecting historical data of the cell to be predicted for the mobile communication network load index. The normalized time series X' = {x1', x2', ..., x m '}:

[0088]

[0089] Where X.min is the minimum element in the X vector, X.max is the maximum element in the X vector, and i = 1, 2, ..., m. Preferably, max = 1 and min = 0, which means that the values ​​of all elements are scaled to between 0 and 1. This can enhance stability for X with very small variance. When the mobile communication network load indicator data approaches 0, the entries in the sparse matrix can be maintained as 0.

[0090] For the normalized time series X'={x1',x2',…,x m '}, perform standardization processing to obtain the standardized time series to be predicted X"={x1",x2",…,x m ”}, the standardization formula is as follows:

[0091]

[0092] where μ and σ are the mean and variance of all elements in the normalized time series X', respectively, and j = 1, 2, ..., m.

[0093] Correspondingly, the sample clean time series of the M cells used for prediction model training is obtained by processing the sample historical data of the mobile communication network load indicators of the M cells collected in the previous N days to obtain the sample original time series using the same preprocessing method as the preset rules specifically defined in the embodiment of the present invention, that is, the preset rules provided in the embodiment of the present invention include the specific steps of data format conversion, outlier removal, missing data filling, normalization and standardization, then the sample historical data of the mobile communication network load indicators of the M cells collected in the previous N days to obtain the sample original time series is also processed by the same specific steps of data format conversion, outlier removal, missing data filling, normalization and standardization provided in the embodiment of the present invention to obtain the sample clean time series of the M cells used for model training. At the same time, the sample cleaning time series of the M cells obtained by the above preset rules need to be divided into a training set and a validation set according to a ratio of 0.75:0.25 for training and validation data of the prediction model respectively. Preferably, M=30K, that is, the sample cleaning time series of 30,000 cells are required to obtain the prediction model. At the same time, the number of elements in the sample cleaning time series is preferably 30×24, that is, the historical data of the previous month is collected, and 24 indicator data are collected every day (collected once every hour). Therefore, the sample cleaning time series composed of the historical data of the rth cell is and the corresponding label y r It is expressed by the following formula:

[0094]

[0095]

[0096] Where r=1,2,…,M, Y is the standard value of the mobile communication network load index at the hour b on the a-th day in the first 30 days of the r-th cell, r To correspond to Tags, is the value of the mobile communication network load index at the hour c on the day of the rth cell.

[0097] The method provided by the embodiment of the present invention converts natural numbers into the logarithmic domain, then removes outliers and fills in missing points on the data, and finally normalizes and standardizes the data. In this way, the data input to the model can be valid data. At the same time, the sample data used for model training is also converted from natural numbers to the logarithmic domain, then outliers are removed and missing points are filled in, and finally the data is normalized and standardized. This allows the network model to better fit data changes. At the same time, it avoids gradient vanishing or gradient exploding in the training network and accelerates the network convergence speed.

[0098] Based on any of the above embodiments, in this method, the outlier detection model is obtained using a One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, specifically including:

[0099] The objective function and constraints of the following optimization problem are constructed to solve the outlier detection model using the One-Class SVM outlier detection algorithm:

[0100]

[0101]

[0102]

[0103] Where a and R are the center and radius of the hypersphere constructed based on n normal sample logarithmic data, R>0, C is the preset penalty coefficient, is the preset slack variable, x i Represents the distance from the i-th mobile communication network load index data point to the center. The sample logarithmic time series of the first N days of M cells contains n normal sample logarithmic data.

[0104] Specifically, due to the huge scale of mobile communication networks and the large number of cells, some abnormal load index values ​​are inevitably generated. In order to avoid the impact of abnormal index values ​​on the performance of the entire system, an outlier detection method is used. The embodiment of the present invention uses One-Class SVM as a monitoring method for outliers, performs statistical analysis based on massive historical data, and obtains a characterization of the entire normal sample group by fitting the normal data of the training data. Optionally, the SVDD (Support Vector Data Description) algorithm is used to solve the One-class SVM to obtain the spherical boundary of the normal data in the algorithm feature space. The algorithm optimization process is the process of minimizing the volume of the hypersphere, and at the same time, the process of minimizing the impact of outliers. It can be expressed as follows:

[0105]

[0106]

[0107]

[0108] Where a and R are the center and radius of the hypersphere constructed based on n normal sample logarithmic data, R>0, C is the preset penalty coefficient, is the preset slack variable, x iRepresents the distance from the i-th mobile communication network load index data point to the center. The sample logarithmic time series of the first N days of M cells contains n normal sample logarithmic data.

[0109] When a new data point is identified using the trained hypersphere, if the data point falls within the hypersphere, it is considered normal data. Otherwise, it is classified as abnormal data and can be removed. After removal, the original position is marked with None to indicate a missing state.

[0110] Based on any of the above embodiments, in this method, the completion model is obtained by training a random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells, specifically including:

[0111] For the tth iteration of training,

[0112] Fitting and Get the fitting matrix and train the current iteration round random forest model Rft. Input Rft, output

[0113] use Update the fitting matrix, s=1, 2, ..., k;

[0114] Where t = 1, 2, …, γ, γ is the preset maximum iteration threshold, k is the number of cells with missing data in the normal logarithmic time series of samples of the first N days of M cells, is the observed value of the sample normal logarithmic time series of the sth cell with missing data, Completion prediction of missing values ​​for the normal logarithmic time series of samples in the sth cell with missing data, is the observed value of the normal logarithmic time series of samples of M cells except the missing data corresponding to the s-th cell in the normal logarithmic time series of other samples except the s-th cell, is the observed value of the missing data corresponding to the sth cell in the normal logarithmic time series of samples of the M cells except the sth cell;

[0115] The random forest model Rfγ obtained by the γth iterative training is the completion model.

[0116] Specifically, the prediction of mobile communication network load indicators is strongly related to the time axis, and the continuity of the time axis is an important prerequisite for ensuring the accuracy of data prediction. However, in the actual load indicator statistics process, it was found that the missing load indicators of a single cell are very common. The main reasons for the missing data include abnormal cell network withdrawal, abnormal elimination of statistical platform data, and system time disorder of the data acquisition platform. In order to avoid the impact of missing time axis data on subsequent network training and prediction, and reduce the prediction accuracy and performance, it is necessary to intelligently supplement it. The embodiment of the present invention targets the continuous variable characteristics of the mobile communication network load indicator data, utilizes the sequence change characteristics of the data set itself, fits the data change law, performs regression prediction, and uses the tree model algorithm-MissForrest to perform data recovery. The specific recovery process is: the sample normal time series set G of M cells is an m*M feature matrix, where m is the number of elements in the sample normal time series of each cell, G represents the input data, and for the sample normal time series G of the sth cell s , s=1,……,k, where k is the number of cells with missing data in the normal logarithmic time series of samples from the first N days of M cells; in order to fill its missing values, we can divide the data into 4 parts:

[0117] is the observed value of the sample normal logarithmic time series of the sth plot with missing data;

[0118] Completion prediction of missing values ​​for the normal logarithmic time series of samples in the sth cell with missing data;

[0119] is the observed value of the normal logarithmic time series of samples of the M cells except the missing data corresponding to the s cell in the normal logarithmic time series of other samples except the s cell;

[0120] is the observed value of the missing data corresponding to the sth cell in the normal logarithmic time series of samples of the M cells except the sth cell;

[0121] Then set the maximum iteration number threshold γ as the algorithm iteration stop condition; is the intermediate value of the completion process, G imp The final completion result for missing values;

[0122] Obtain the completion model through the following algorithm:

[0123] Missing values ​​are randomly initialized;

[0124] The columns of k←G are arranged in ascending order of the amount of missing data;

[0125]

[0126]

[0127] The final output G imp The final completion result is the random forest model Rfγ generated by the last iteration as the completion model.

[0128] Based on any of the above embodiments, in this method, the network structure used in the prediction model training is a BiGRU network with an Attention mechanism added, specifically including:

[0129] The prediction model is trained using a 1-layer BiGRU network layer and the output h of the 1-layer BiGRU network layer is converted into l And the output Y of all BiGRU network layers is input into the Attention mechanism layer;

[0130] The Attention mechanism layer performs the following operations:

[0131] Based on h l and Y determine the attention feature score of each BiGRU network layer at each moment;

[0132] Determine the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment;

[0133] Determine a weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers;

[0134] Based on the weighted expression of all BiGRU network layer outputs and h l Determine the output of the Attention mechanism layer.

[0135] Specifically, the BiGRU network is introduced into the training network used in the prediction model training, and the Attention mechanism is added. Figure 2 A schematic diagram of a network structure in which a multi-layer BiGRU network is superimposed is provided in an embodiment of the present invention. Figure 2 As shown, Figure 2 The BiGRU network layer in the total number of L layers ( Figure 2 The "L Layers" on the left) contains T features for each BiGRU network layer, because the input X = (x1, x2, x3, ..., x T), there are T elements in the sequence X. When the training network of the prediction model provided by the embodiment of the present invention uses a multi-layer BiGRU network layer structure, X corresponds to the sample clean time series provided by the embodiment of the present invention, T corresponds to the number of elements in the sample clean time series provided by the embodiment of the present invention, and the output Y of all BiGRU network layers is (y1, y2, y3, ..., y T ). Then the BiGRU network is combined with the Attention mechanism. Specifically, based on the output h of the first BiGRU network layer, l The attention feature scores of each BiGRU network layer at each moment are determined based on the attention feature scores of each BiGRU network layer at each moment; the weight values ​​corresponding to the current moment of each BiGRU network layer are determined based on the weight values ​​corresponding to the current moment of each BiGRU network layer; the weighted expression of the outputs of all BiGRU network layers is determined based on the weighted expression of the outputs of all BiGRU network layers and h l Determine the output of the Attention mechanism layer, where l represents the number of BiGRU network layers.

[0136] Based on any of the above embodiments, in the method, the h-based l and Y determine the attention feature scores of each BiGRU network layer at each moment, including:

[0137] The attention feature score MQ of each BiGRU network layer at each moment is determined by the following formula:

[0138] MQ=tanl(W y ×Y+(W h ×h l ×e l ))

[0139] Among them, l is the number of BiGRU network layers, h l ∈R 1×l The first layer is the output of the BiGRU network layer, MQ∈R F×l , F is the number of elements in the cleaning time series of each cell sample, W y ∈R F×F , Wh∈R F×F , e l ∈R F And W y 、W h and e l are randomly generated trainable matrices, matrices, and vectors, Y∈R F×l And it is the output of all BiGRU network layers;

[0140] Correspondingly, determining the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment specifically includes:

[0141] The weight value aa corresponding to each BiGRU network layer at the current moment is determined by the following formula:

[0142] aa=softmax(ω T ×MQ)

[0143] Where aa∈R l ,ω∈R F And ω is a randomly generated trainable vector, and softmax(·) is a normalized exponential function;

[0144] The determining of the weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers specifically includes:

[0145] The weighted expression rr of all BiGRU network layer outputs is determined by the following formula:

[0146] rr=Yaa T

[0147] Where rr∈R F ;

[0148] The weighted expression based on the output of all BiGRU network layers and h l Determine the output of the Attention mechanism layer, including:

[0149] The output h of the Attention mechanism layer is determined by the following formula * :

[0150] h * =tanl(W p ×rr+W x ×h l )

[0151] Among them, W p ∈R F×F , W x ∈R F×F , and W p and W x All are randomly generated trainable matrices.

[0152] Specifically, F is the number of elements in the clean time series of each cell sample, which is also equivalent to the dimension of the features extracted by each BiGRU network layer. l is the number of layers in the BiGRU network layer, that is, the hidden layer dimension. W y 、W h 、Wp 、W x , ω and e l are randomly generated trainable matrices, matrices, matrices, matrices, vectors, and vectors, respectively, Y(Y∈R F×l ) is the output of all BiGRU network layers, that is, the output of all hidden layers of the BiGRU network, h l (h l ∈R 1×l ) is the output of the first BiGRU network layer, that is, the last hidden layer output. MQ represents the attention feature score of each BiGRU network layer at each moment. It should be noted here that each moment of a certain BiGRU network layer corresponds to each feature of the said BiGRU network layer. The larger the value of the attention feature score, the greater the corresponding allocated attention; aa(aa∈R l ) represents the weight value corresponding to each BiGRU network layer at the current moment; rr(rr∈R F ) is the weighted expression of the output of all BiGRU network layers, that is, the attention expression inherits the content involved in the previous step of the network model; h * (h * ∈R F ) is the final output feature, which is composed of the nonlinear combination of the attention expression feature at the last moment and the final output of the hidden layer feature. * Input into the subsequent neural network.

[0153] Based on any of the above embodiments, an embodiment of the present invention provides an Attention-BiGRU network model for training a prediction model. Figure 3 A schematic diagram of the structure of the Attention-BiGRU network model for training a prediction model provided by an embodiment of the present invention is shown in FIG. Figure 3As shown, here, the preferred number of BiGRU network layers is 128. The outputs of all BiGRU network layers enter the Attention mechanism layer for processing. The output then continues to the Flatten layer to reduce the multidimensional data to one dimension. The output then passes through the Dense layer and finally enters the Activation layer for processing using the activation function to obtain the output. The load index data of the sample cells used for prediction model training in the previous 30 days are extracted, and the load index data at the hourly point is extracted every day. That is, the sample cleaning time series of each cell includes 30×24 elements, and the load index value at the hourly point of each hour of the 24 hours of the day is used as the label. 30,000 representative cells within the preset range are selected to form the model training set and validation data set, which are divided into training set and validation set at a ratio of 0.75:0.25. The MSE (mean square error) method was used for training for 8000 epochs. The momentum algorithm optimizer was used for model training. The learning rate was set to 0.002, momentum = 0.8, and global step = 200. The network model training process was completed through cyclic iterations. The training was stopped when the preset maximum number of iterations threshold was reached.

[0154] Based on any of the above embodiments, an embodiment of the present invention provides a cell network load index prediction device based on an attention mechanism learning network. Figure 4 A schematic diagram of the structure of a cell network load index prediction device for an attention mechanism learning network provided by an embodiment of the present invention. Figure 4 As shown, the apparatus includes a pre-processing unit 410 and a prediction unit 420, wherein:

[0155] The preprocessing unit 410 is configured to collect historical data of the mobile communication network load index to be predicted for the previous N days of the cell of the mobile communication network load index to be predicted to obtain an original time series to be predicted, and process the original time series to be predicted according to a preset rule to obtain a clean time series to be predicted, where N is a positive integer;

[0156] The prediction unit 420 is used to input the cleaning time series to be predicted into a prediction model and output the predicted mobile communication network load index of the cell on that day; wherein the prediction model is obtained after training based on the sample cleaning time series of M cells and the mobile communication network load index labels corresponding to the M cells on that day, and the sample cleaning time series of the M cells are obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load indicators of the M cells in the previous N days according to the preset rules; the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0157] The device provided by the embodiment of the present invention collects historical data of the mobile communication network load index to be predicted for the previous N days of the cell whose mobile communication network load index is to be predicted to obtain an original time series to be predicted, and processes the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; the clean time series to be predicted is input into a prediction model, and the predicted mobile communication network load index of the cell on that day is output; wherein the prediction model is obtained after training based on sample clean time series of M cells and mobile communication network load index labels of the M cells on that day, and the sample clean time series of the M cells is obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load index of the M cells in the previous N days according to the preset rules, and the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added. Because the prediction model for mobile communication network load index prediction is dynamically generated through deep learning, and the BiGRU network used for model training incorporates an attention mechanism, which is used during training to screen important features and filter out low-value features, the prediction model's ability to express mobile communication network load index data collected over a long period of time is improved. Through iterative training, the model can more accurately fit the changing trends of mobile communication network load index data. Therefore, the apparatus provided by the embodiments of the present invention achieves improved accuracy in predicting cell mobile communication network load indexes.

[0158] Based on any of the above embodiments, in the device, the step of processing the original time series to be predicted according to a preset rule to obtain the cleaning time series to be predicted specifically includes:

[0159] Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted;

[0160] Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted;

[0161] Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted;

[0162] Normalizing and standardizing the to-be-predicted completion time series to obtain a to-be-predicted cleaning time series;

[0163] The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by training the random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells.

[0164] The device provided by the embodiment of the present invention converts natural numbers into the logarithmic domain, then removes outliers and fills in missing points on the data, and finally normalizes and standardizes the data. In this way, the data input to the model can be valid data. At the same time, the sample data used for model training is also converted from natural numbers to the logarithmic domain, and then outliers are removed and missing points are filled in. Finally, the data is normalized and standardized, which allows the network model to better fit the data changes. At the same time, it avoids gradient vanishing or gradient explosion in the training network and accelerates the network convergence speed.

[0165] Based on any of the above embodiments, in the device, the outlier detection model is obtained using a One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, specifically including:

[0166] The objective function and constraints of the following optimization problem are constructed to solve the outlier detection model using the One-Class SVM outlier detection algorithm:

[0167]

[0168]

[0169]

[0170] Where a and R are the center and radius of the hypersphere constructed based on n normal sample logarithmic data, R>0, C is the preset penalty coefficient, is the preset slack variable, x i Represents the distance from the i-th mobile communication network load index data point to the center. The sample logarithmic time series of the first N days of M cells contains n normal sample logarithmic data.

[0171] Based on any of the above embodiments, in the device, the completion model is obtained by training a random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells, specifically including:

[0172] For the tth iteration of training,

[0173] Fitting and Get the fitting matrix and train the current iteration round random forest model Rf t ,Will Input Rf t , output

[0174] use Update the fitting matrix, s=1, 2, ..., k;

[0175] Where t = 1, 2, …, γ, γ is the preset maximum iteration threshold, k is the number of cells with missing data in the normal logarithmic time series of samples of the first N days of M cells, is the observed value of the sample normal logarithmic time series of the sth cell with missing data, Completion prediction of missing values ​​for the normal logarithmic time series of samples in the sth cell with missing data, is the observed value of the normal logarithmic time series of samples of M cells except the missing data corresponding to the s-th cell in the normal logarithmic time series of other samples except the s-th cell, is the observed value of the missing data corresponding to the sth cell in the normal logarithmic time series of samples of the M cells except the sth cell;

[0176] The random forest model Rfγ obtained by the γth iterative training is the completion model.

[0177] Based on any of the above embodiments, in the device, the network structure used in the prediction model training is a BiGRU network with an Attention mechanism added, specifically including:

[0178] The prediction model is trained using a 1-layer BiGRU network layer and the output h of the 1-layer BiGRU network layer is converted into l And the output Y of all BiGRU network layers is input into the Attention mechanism layer;

[0179] The Attention mechanism layer performs the following operations:

[0180] Based on h l and Y determine the attention feature score of each BiGRU network layer at each moment;

[0181] Determine the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment;

[0182] Determine a weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers;

[0183] Based on the weighted expression of all BiGRU network layer outputs and h l Determine the output of the Attention mechanism layer.

[0184] Based on any of the above embodiments, in the device, the h l and Y determine the attention feature scores of each BiGRU network layer at each moment, including:

[0185] The attention feature score MQ of each BiGRU network layer at each moment is determined by the following formula:

[0186] MQ=tanl(W y ×Y+(W h ×h l ×e l ))

[0187] Among them, l is the number of BiGRU network layers, h l ∈R 1×l The first layer is the output of the BiGRU network layer, MQ∈R F×l , F is the number of elements in the cleaning time series of each cell sample, W y ∈R F×F , Wh∈R F×F , e l ∈R F And W y 、W h and e l are randomly generated trainable matrices, matrices, and vectors, Y∈R F×l And it is the output of all BiGRU network layers;

[0188] Correspondingly, determining the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment specifically includes:

[0189] The weight value aa corresponding to each BiGRU network layer at the current moment is determined by the following formula:

[0190] aa=softmax(ω T ×MQ)

[0191] Where aa∈R l ,ω∈R F And ω is a randomly generated trainable vector, and softmax(·) is a normalized exponential function;

[0192] The determining of the weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers specifically includes:

[0193] The weighted expression rr of all BiGRU network layer outputs is determined by the following formula:

[0194] rr=Yaa T

[0195] Where rr∈R F ;

[0196] The weighted expression based on the output of all BiGRU network layers and hl Determine the output of the Attention mechanism layer, including:

[0197] The output h of the Attention mechanism layer is determined by the following formula * :

[0198] h * =tanl(W p ×rr+W x ×h l )

[0199] Among them, W p ∈R F×F , W x ∈R F×F , and W p and W x All are randomly generated trainable matrices.

[0200] Figure 5 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention is shown in FIG. Figure 5 As shown, the electronic device may include: a processor (processor) 501, a communication interface (Communications Interface) 502, a memory (memory) 503 and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504. The processor 501 can call a computer program stored in the memory 503 and executable on the processor 501 to execute the cell network load index prediction method based on the attention mechanism learning network provided in the above-mentioned embodiments, for example, including: collecting historical data of the mobile communication network load index to be predicted for the cell of the mobile communication network load index to be predicted in the previous N days to obtain the original time series to be predicted, and processing the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; inputting the clean time series to be predicted into the prediction model, and outputting the predicted mobile communication network load index of the cell on that day; wherein the prediction model is obtained after training based on sample clean time series of M cells and mobile communication network load index labels corresponding to the M cells on that day, and the sample clean time series of the M cells is obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load index of the M cells in the previous N days according to the preset rules; the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0201] In addition, the logic instructions in the above-mentioned memory 503 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0202] An embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the cell network load index prediction method based on the attention mechanism learning network provided in the above-mentioned embodiments, for example, including: collecting historical data of the mobile communication network load index to be predicted for the cell of the mobile communication network load index to be predicted in the previous N days to obtain an original time series to be predicted, processing the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; inputting the clean time series to be predicted into a prediction model, and outputting the predicted mobile communication network load index of the cell on that day; wherein the prediction model is obtained after training based on sample clean time series of M cells and mobile communication network load index labels corresponding to the M cells on that day, and the sample clean time series of the M cells is obtained by processing the sample original time series obtained by collecting sample historical data of the mobile communication network load index of the M cells in the previous N days according to the preset rules, and the network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer.

[0203] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0204] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for predicting cell network load indicators based on an attention mechanism learning network, characterized in that: include: Collecting historical data of the mobile communication network load index to be predicted for the previous N days of the cell for the mobile communication network load index to be predicted to obtain an original time series to be predicted, and processing the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, where N is a positive integer; the preset rules include data format conversion, outlier removal based on an outlier detection model, missing data completion based on a completion model, normalization, and standardization; The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by using the random forest classifier training based on the normal logarithmic time series of samples from the previous N days of M cells; Inputting the cleaning time series to be predicted into the prediction model, and outputting the predicted mobile communication network load index of the cell on that day; The prediction model is obtained through training based on the sample cleaning time series of M cells and the mobile communication network load index labels corresponding to the M cells on the same day. The sample cleaning time series of the M cells is obtained by processing the sample original time series obtained by collecting the sample historical data of the mobile communication network load index of the M cells in the previous N days and performing the preset rule processing. The network structure used in the training of the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer. The network structure used in the prediction model training is a BiGRU network with an Attention mechanism, which specifically includes: The prediction model is trained using a 1-layer BiGRU network layer and the output h of the 1-layer BiGRU network layer is converted into l And the output Y of all BiGRU network layers is input into the Attention mechanism layer; The Attention mechanism layer performs the following operations: Based on h l and Y determine the attention feature score of each BiGRU network layer at each moment; Determine the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment; Determine a weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers; Based on the weighted expression of all BiGRU network layer outputs and h l Determine the output of the Attention mechanism layer; The h-based l and Y determine the attention feature scores of each BiGRU network layer at each moment, including: The attention feature score MQ of each BiGRU network layer at each moment is determined by the following formula: ； Among them, l is the number of layers of BiGRU network layer, The first is the output of the first BiGRU network layer, MQ , F is the number of elements in the cleaning time series of each cell sample, , and 、 and are randomly generated trainable matrices, matrices, and vectors, respectively, Y And it is the output of all BiGRU network layers; Correspondingly, determining the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment specifically includes: The weight value aa corresponding to each BiGRU network layer at the current moment is determined by the following formula: ； Among them, aa , and is a randomly generated trainable vector, softmax is the normalized exponential function; The determining of the weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers specifically includes: The weighted expression rr of all BiGRU network layer outputs is determined by the following formula: ； Among them, rr ; The weighted expression based on the output of all BiGRU network layers and h l Determine the output of the Attention mechanism layer, including: The output of the Attention mechanism layer is determined by the following formula : ； in, , ,and and All are randomly generated trainable matrices.

2. The method for predicting cell network load indicators based on an attention mechanism learning network according to claim 1, characterized in that: The processing of the original time series to be predicted with preset rules to obtain the cleaning time series to be predicted specifically includes: Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted; Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted; Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted; The to-be-predicted completion time series is normalized and standardized to obtain the to-be-predicted cleaning time series.

3. The method for predicting cell network load indicators based on attention mechanism learning network according to claim 2 is characterized in that: The outlier detection model is based on the logarithmic time series of samples from the previous N days of M cells using the One-Class SVM outlier detection algorithm. Specifically, it includes: The objective function and constraints of the following optimization problem are constructed to solve the outlier detection model using the One-Class SVM outlier detection algorithm: ； ； ； Where a and R are the center and radius of the hypersphere constructed based on n normal sample logarithmic data, R>0, is the preset penalty coefficient, is the preset slack variable, Representative The distance from a mobile communication network load index data point to the center, and the sample logarithmic time series of the first N days of M cells contain n normal sample logarithmic data.

4. The method for predicting cell network load indicators based on an attention mechanism learning network according to claim 2, characterized in that: The completion model is trained using a random forest classifier based on the normal logarithmic time series of samples from the previous N days of M cells, specifically including: For the tth iteration of training, Fitting and Get the fitting matrix and train the current iteration round random forest model Rf t , Will Input Rf t , output , use Update the fitting matrix, s=1,2,…,k; Where t=1,2,…,γ,γ is the preset maximum iteration threshold, k is the number of cells with missing data in the normal logarithmic time series of samples of the first N days of M cells, is the observed value of the sample normal logarithmic time series of the sth cell with missing data, Completion prediction of missing values ​​for the normal logarithmic time series of samples in the sth cell with missing data, is the observed value of the normal logarithmic time series of samples of M cells except the missing data corresponding to the s-th cell in the normal logarithmic time series of other samples except the s-th cell, is the observed value of the missing data corresponding to the sth cell in the normal logarithmic time series of samples of the M cells except the sth cell; The random forest model Rfγ obtained by the γth iterative training is the completion model.

5. A cell network load index prediction device based on an attention mechanism learning network, characterized in that: include: A preprocessing unit is configured to collect historical data of the mobile communication network load index to be predicted for the previous N days of the cell for which the mobile communication network load index is to be predicted, obtain an original time series to be predicted, and process the original time series to be predicted according to preset rules to obtain a clean time series to be predicted, wherein N is a positive integer; the preset rules include data format conversion, outlier removal based on an outlier detection model, missing data completion based on a completion model, normalization, and standardization; The outlier detection model is obtained by using the One-Class SVM outlier detection algorithm based on the logarithmic time series of samples from the previous N days of M cells, and the completion model is obtained by using the random forest classifier training based on the normal logarithmic time series of samples from the previous N days of M cells; A prediction unit is configured to input the to-be-predicted cleaning time series into a prediction model and output a predicted mobile communication network load index for the cell on that day; wherein the prediction model is obtained through training based on sample cleaning time series of M cells and labels of mobile communication network load indexes corresponding to the M cells on that day; the sample cleaning time series of the M cells are obtained by processing the sample original time series obtained by collecting sample historical data of mobile communication network load indicators of the M cells over the previous N days and subjecting them to the preset rules; the network structure used in training the prediction model is a BiGRU network with an Attention mechanism added, and M is a positive integer; The network structure used in the prediction model training is a BiGRU network with an Attention mechanism, which specifically includes: The prediction model is trained using a 1-layer BiGRU network layer and the output h of the 1-layer BiGRU network layer is converted into l And the output Y of all BiGRU network layers is input into the Attention mechanism layer; The Attention mechanism layer performs the following operations: Based on h l and Y determine the attention feature score of each BiGRU network layer at each moment; Determine the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment; Determine a weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers; Based on the weighted expression of all BiGRU network layer outputs and h l Determine the output of the Attention mechanism layer; The h-based l and Y determine the attention feature scores of each BiGRU network layer at each moment, including: The attention feature score MQ of each BiGRU network layer at each moment is determined by the following formula: ； Among them, l is the number of layers of BiGRU network layer, The first is the output of the first BiGRU network layer, MQ , F is the number of elements in the cleaning time series of each cell sample, , and 、 and are randomly generated trainable matrices, matrices, and vectors, respectively, Y And it is the output of all BiGRU network layers; Correspondingly, determining the weight value corresponding to each BiGRU network layer at the current moment based on the attention feature score of each BiGRU network layer at each moment specifically includes: The weight value aa corresponding to each BiGRU network layer at the current moment is determined by the following formula: ； Among them, aa , and is a randomly generated trainable vector, softmax is the normalized exponential function; The determining of the weighted expression of the outputs of all BiGRU network layers based on the weight values ​​corresponding to the current moments of the BiGRU network layers specifically includes: The weighted expression rr of all BiGRU network layer outputs is determined by the following formula: ； Among them, rr ; The weighted expression based on the output of all BiGRU network layers and h l Determine the output of the Attention mechanism layer, including: The output of the Attention mechanism layer is determined by the following formula : ； in, , ,and and All are randomly generated trainable matrices.

6. The cell network load index prediction device based on the attention mechanism learning network according to claim 5 is characterized in that: The processing of the original time series to be predicted with preset rules to obtain the cleaning time series to be predicted specifically includes: Converting the natural numbers in the original time series to be predicted into logarithms to obtain a logarithmic time series to be predicted; Input the logarithmic time series to be predicted into the outlier detection model, and output the normal logarithmic time series to be predicted; Inputting the normal logarithmic time series to be predicted into the completion model, and outputting the completed time series to be predicted; The to-be-predicted completion time series is normalized and standardized to obtain the to-be-predicted cleaning time series.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the cell network load index prediction method based on the attention mechanism learning network as described in any one of claims 1 to 4 are implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting a cell network load index based on an attention mechanism learning network as described in any one of claims 1 to 4 are implemented.

Images